Rational Nonmonotonic Reasoning

Transcription

1 CARL M. KADE Rational Nonmonotonic Reasoning Department of Computer Science, University of illinois, Urbana, L U.S.A. Abstract. Nonmonotonic reasoning is a pattern of reasoning that allows an agent to make retract (tentative) conclusions from inconclusive evidence. This paper gives a possible-worlds interpretation of the nonmonotonic reasoning problem based on stard decision theory the emerging probability logic. The system's central principle is that a tentative conclusion is a decision to make a bet, not an assertion of fact The system is rational, as sound as the proof theory of its underlying probability logic. 1 ntroduction Background The ability to make a tentative conclusion from inconclusive evidence then retract that conclusion if discrediting evidence is later acquired is characteristic of intelligence. Many theories of nonmonotonic reasoning, default reasoning, inductive logic try to model this behavior. But, because these theories do not fully consider the consequences context of adopting a conclusion, they can behave irrationally. As the expression "Don't bet your life on it" illustrates, hasty conclusions can be disastrous. Decision theory the emerging probability logic offer a straightforward possible-worlds interpretation of idealized nonmonotonic reasoning that adapts to the context of the reasoning problem. 1.1 The Bird Example The bird example demonstrates nonmonotonic reasoning. The nonmonotonic reasoner is told: "Tweety is a bird." t is asked: "Can Tweety fly?" According to the literature, it answers: "Yes, Tweety can fly." Now it is given another fact: "Tweety is a Penguin." Now, how should it answer: "Can Tweety Fly?" According to the literature, it answers: "No, Tweety can't fly." The addition of new facts caused the nonmonotonic reasoner to retract a conclusion. n conventional logic, new facts never cause an agent to retract a conclusion; thus, conclusions grow monotonically with information. Consequently, conventional logic must be more conservative in making conclusions than nonmonotonic reasoning. 1.2 Current Nonmonotonic Reasoning Systems Logic probability are two popular approaches to nonmonotonic reasoning Logic-Based Systems Conceptually, a logic-based nonmonotonic reasoner makes its conclusions in two steps. First, from its defaults knowledge it creates deductively-closed belief sets called extensions (Ginsberg, 1987). n the Tweety example, with default bird(x) fly(x) the knowledge that bird(tweety) penguin(x) -,fly(x),it creates thesingle extension{bird(tweety), penguin(x) -,fly(x),fly(tweety)}. The second part of making conclusions is the application of an acceptance rule. The rule is a procedure that tells the nonmonotonic reasoner what conclusions to draw from the extensions. When a nonmonotonic system creates a single extension, the acceptance rule is simple: accept (or conclude) a sentences if it is in the (deductive closure of the) extension. 197

2 Continuing the example, if the system is told that penguin (Tweety ), the default rule cannot fire because its consequent contradicts the fact--, fly(tweety) (from penguin (Tweety) penguin (x) _, --,fly (x )). Thus, it now concludes that Tweety cannot fly because it now has the single extension: {bird(tweety), penguin(x) -7 -,fly(x), -,fly(tweety)}. The acceptance rules are more interesting when defaults conflict. For example, suppose the reasoner knows that Quaker(Nixon) Republican(Nixon) that it has defaults Quaker(x) -7 pacifi.st(x) Republican(x) _,..., pacifist(x) (Reiter, Criscuolo, 1981). Now, rather than one extension there are two (Moore, 1985): {Quaker(Nixon ), Republican (Nixon), pacifist(nixon )} {Quaker(Nixon ), Republican(Nixon ), --, pacifist(nixon )}. Reiter's system accepts a sentence (tentatively concludes that it is so) if the sentence is true in any extension (Reiter, 1980). Consequently, it accepts the sentence pacijist(nixon) the sentence..., pacijist(nixon). McDermott (1982) calls this acceptance rule brave. Circumscription offers a different acceptance rule: it accepts a sentence if the sentence is true in every extension. Thus, circumscription makes no conclusion about pacifist(nixon) no conclusion about..., pacijist(nixon). McDennott calls this acceptance rule cautious. These two acceptance rules will be called modal. f the set of extensions is treated as the set of possible worlds in an S5 modal structure, Reiter's acceptance rule corresponds closely to the model operator o (possibility) circumscription's rule corresponds closely to the 0 (necessity) operator (Dowty, Wall, Peters, 1981). Many systems forgo the modal approach by evaluating all sentences in a particular extension. The extension can be chosen arbitrarily (Doyle, 1983), according to some hierarchy of defaults (or priorities of circumscription (McCarthy, 1980; Lifschitz, 1987)), by choosing the most probable extension (Pearl, 1987), or by some other method. n these pick-an-extension systems the truth value of a complex sentence is a function of the truth values of the literals that make up that sentence. This is not true in the modal systems. For example, Reiter's modal system does not accept the sentence pacifist(nixon) 1\..., pacifist(nixon), but it does accept the literal pacifist(nixon) the literal..., pacifi.st(nixon). Likewise, circumscription accepts pacifist(nixon) v --, paclfzst(nixon ), but not pacifi.st(nixon ), not-, pacifist(nixon ) Probability Based Systems Like logic-based systems, probability-based nonmonotonic reasoning systems (conceptually) make conclusions in two steps. First, they assign some degree of belief to a sentence. The degree might be a probability measure, a certainty factor, a Zadeh possibility degree, or a Dempster-Shafer interval. Second, using the degree of belief, an acceptance rule decides whether to make the sentence a conclusion. (Some A systems (Rich, 1983; Ginsberg, 1984; Farreny & Prade, 1986; Dubois & Prade, 1988) ignore the acceptance step). The most obvious acceptance rule is: AcceptS if, only if, P(S) > 1-, for some small e. P(S) is the probability of sentences. The rule says: if a sentence is certain enough, say 99.99% certain, conclude that it is so. 1.3 Scope The body of this paper concerns the desired properties of a nonmonotonic reasoning system the specification of such a system. Before starting, it is appropriate to discuss what will will not be done. First, this paper will take no st in the long-running debate as to the practicality of probabilities:., 198

3 The information necessary to assign numerical probabilities is not ordinarily available. Therefore, a formalism that required numerical probabilities would be epistemologically inadequate. (McCarthy & Hayes,J969) And on the other h:... many extensions of logic, such as' default logic' are better understood in a probabilistic framework. (Cheeseman, 1985) Second, this paper will present a fonnal account of nonmonotonic reasoning. ts approach may superficially appear less fonnal than approaches that invent a new fonnal system provide proofs of its properties. The approach taken here unifies (Ginsberg, 1987) nonmonotonic reasoning with a previously developed fonnal system. This approach may legitimately rely on others to develop the formal properties of the formal system. Third, this paper will not prove that this nonmonotonic system is correct or optimal because there is no stard to which it can be compared. All that will be argued is that the system offers useful insight into interesting problems, that it make only reasonable assumptions, that it is internally consistent. Now that the general scope of the desired nonmonotonic system is delimited, its particular properties are specified. 2 Desired Properties The desired properties of soundness rationality are described justified in tum. 2.1 Soundness One test of a nonmonotonic system's soundness is Kyburg' s lottery paradox (Kyburg, 1970). The paradox shows that nonmonotonic systems that allow chaining (that is, introduction, modus ponens, so on) across extensions (for example (Ginsberg, 1985; Yager, 1987)) are unsound; that is, they allow inconsistent conclusions. Kyburg's example was this: suppose there is a lottery with 1 00 tickets. A reasoning system knows that exactly one ticket will win. Suppose it has a default that says that (unless it knows to the contrary) it should assume that loser(ticket_#1). Similarly, suppose it has a default that says loser(ticket_#2) so on for all the tickets. These defaults the fact that exactly one ticket will win lead to extensions. n each extension a different ticket wins. f the system is brave with respect to the literals, its set of conclusions will be loser(ticket _# 1 ), loser(ticket _#2),..., loser( ticket_# And if chaining is allowed across extensions, then by introduction it will conclude that loser(ticket_#) A loser(ticket_#2) A A loser(ticket_#oooo). That is, it will conclude that no ticket will win. This conclusion contradicts its knowledge that one ticket will win. Two alternatives to chaining across extensions were mentioned earlier: pick -an-extension systems, which chain within a single extension, modal systems, which use many extensions, but allow no chaining. As a third alternative, a system may be particularly cautious about accepting literals; it may accept only the deductive closure of the literals true in every extension. n the lottery example, because no literal is true in every extension, such a system would accept nothing; thus it avoids all the problems ( benefits) of nonmonotonic reasoning. 2.2 Rationality The other property to consider is rationality. Rationality is best understood with an example Russian Roulette Example Consider this problem in nonmonotonic reasoning: 199

4 A revolver is loaded with 1 bullet (it has 5 empty chambers), the cylinder is spun. With these stakes: f correct, the system wins $1. f wrong, the system loses $1. would a nonmonotonic system take the bet that the gun will not fire? The answer is yes; most systems would take the bet, because the gun is unlikely to fire. Now consider a second scenario:. bullot, t11 qliltdlr.,. Again the revolver is loaded with exactly 1 bullet the cylinder is spun. With these new stakes: f correct, the system wins $1. f wrong, the system loses its life. would a nonmonotonic system take the bet that the gun will not fire? Most nonmonotonic systems would take the bet -- because, the gun is still unlikely to fire. Obviously something is wrong with this reasoning; in these two scenarios the uncertainty is the same, yet it is not rationa/ 1 to draw the same conclusion. How is the Russian roulette problem solved? Decision theory gives the answer: compare the probability of the sentence to the breakeven probability detennined by the payoff. n the first scenario, the breakeven probability is 11(1 + 1) = 0.5, because P(gun_will_notJire ll_bullet spun) is greater than 0.5, the system should conclude that the gun will not fire. ln the second scenario the breakeven probability is lim 1 = 1. ;-.-l+ The system should ignore a better-than-even probability refuse to bet its life on the proposition that the gun will not fire. 1 Though the exact meaning of the tenn rational is the subject of debate, most authors agree that a rational agent maximizes its expected utility (Savage, 1954; Ferguson, 1967) avoids Dutch books (Kyburg & Smokier 1964 ). 200

5 2.2.2 The Need for Payoff Probability The example shows that rational nonmonotonic systems must consider payoff act accordingly. Systems that fail to consider payoff will in general miss opportunities (the first scenario) or suffer disastrous consequences (the second). How should payoff be measured? f it is assumed there is a total preference ordering among payoffs that the preference ordering of bets is based on expected payoff, utility theory is the answer (Ferguson, 1967; Savage, 1954). Furthennore, if these axioms of utility theory are accepted, it follows that a rational nonmonotonic system must be probabilistic, for without probability, rational choices cannot be detennined. Decision theory allows us to have a nonmonotonic reasoner that automatically adjusts its degree of valor. Depending on the circumstances, it can be cautious like a circumscription system, brave like Reiter's system, or somewhere in between. Consideration of payoff is opposed by some. Pearl (1987) asserts that in day-to-day life the overhead of considering payoff overwhelms the benefit of such deliberations. Others disagree (Levi 1980). n an idealized system there is no debate -- computation is free optimal perfonnance is expected -- so payoff must be considered. 3 Specification Given these desired properties, a description of a rational nonmonotonic system is now straightforward. The nonmonotonic system is based on decision theory probability. t is modal does not allow chaining. 3.1 Probability The first step is the specification of a language, model theory, semantics. One possibility is Nilsson's probabilistic logic (Nilsson, 1986). The intuition behind probabilistic logic is that a particular sentence in a particular world/extension is either true or false. But, because agents do not know which world they are in, they must look at the set of worlds consistent with their knowledge. Agents find the probability of a sentence by summing the probability of each world in which the sentence is true. f the set of literals is finite, the set of all possible worlds their probabilities is expressible as a finite joint probability density. For example the prior density for the bird example might be Can Fly Can't Fly Not a Penguin Not a Bird A Penguin 0 0 Not a Penguin A Bird A Penguin Although probabilistic logic gives the desired model theory, ' it lacks expressiveness, so Frisch & Haddawy's modal probability logic (Frisch & Haddawy, forthcoming), an extended probability logic, is used, making it possible to say things such as P (gun_ will_not Jire _ bullet spun) > 0.5 in the object language. Because the nonmonotonic reasoning system is embedded in a probability logic, it is as sound ( complete) as that probability logic. A proof theory for the propositional probability logic is trivial: simply enumerate the finite possible-worlds add their probabilities. This proof theory is sufficient for the examples considered in this paper. A proof theory for predicate probability logic is still emerging. 201

6 3.2 Acceptance The next step is the specification of the acceptance rule. A sentence S with respect to some evidence e a breakeven probability b is accepted exactly if P(Sie) >b. This rule comes from decision theory. The key to this system, therefore, is that a tentative conclusion is an assertion about the desirability of a bet, not a direct assertion about a sentence. For example, where other systems might conclude fly (Tweety ), this system concludes P ifly(tweety) 1 bird (Tweety )) > 0.6. f later it learns that penguin (Tweety ), it concludes -., P(fly(Tweety) bird(tweety)&penguin(tweety)) > Chaining to take The final part of the specification is a prohibition against chaining conclusions. The system is not allowed deduce P(loser(ticket_#J) _ticket_lottery) >a P(loser(ticket_#2) lloooo_ticket_lottery) > a P(loser(ticket_#l)" loser(ticket_#2) _ticket_lottery) >a. The prohibition is justified because loser(ticket_#1) loser(ticket_#2) might seldom be true in the same world, so, any chaining crosses extensions/worlds. Actually, no explicit prohibition is needed because the semantics (from probability logic) already prohibit chaining across worlds. Also, because sentences such as P(loser(ticket_#l)" loser(ticket_ #2) 110QOO_ticket_lottery) >a, can be evaluated directly, no chaining needed. 3.4 Tweety Revisited Here is how this system renders the bird example. Suppose 1) that the prior joint probability density is Not a Penguin A Penguin 0 Not a Penguin A Penguin 0 2) that the payoff is Can Fly The system says can_jly(tweety). f right, it wins $1.00. f wrong, it loses $1.50. Can't Fly Not a Bird A Bird (the breakeven probabilities are , respectively. The system says -, can_jly(tweety). f right, it wins $1.50. f wrong, it loses $ ) that the reasoner is told bird(tweety). The probabilities calculated from the density are: P(can_jly(Tweety) bird(tweety) = =

7 P(-, canjly(tweety) bird(tweety)) = "" Because 0.82 is greater that 0.6, the system concludes P(can_fly(Tweety) bird(tweety)) > 0.6; because 0.18 is not greater than 0.4 it also concludes -, P(-, can_fly(tweety) bird(tweety)) > 0.4. Or in English: given that Tweety is a bird, it is rational to bet Tweety can fly ( not to bet that Tweety cannot fly). Now suppose that the system is told that Tweety is a penguin. From the probability density the system computes that P(can_fly(Tweety) bird(tweety) 1\ penguin(tweety)) = P(-, can_fly(tweety) bird(tweety) 1\ penguin(tweety)) = t concludes = 1. -, P(can_fly(Tweety) bird(tweety) 1\ penguin(tweety)) > 0.6 P (-, can _fly (Tweety) bird (Tweety) 1\ penguin (Tweety)) > 0.4. That is, given what it knows, it bets that Tweety cannot fly ( it does not bet that Tweety can fly). So the system hles the bird example properly. 4 Conclusion A nonmonotonic reasoning system should be sound rational. The Russian roulette problem shows the dangers of irrational reasoning. A system that does not evaluate probabilities consider payoff may miss opportunities stumble into disasters. With decision theory, a system is automatically as courageous as circumstances warrant. n the extremes, it can be as cautious as a circumscription system or as brave as Reiter's system. Probability logic gives a decision-theory-based nonmonotonic reasoning system a rich possible-worlds semantics. The form of the conclusions is or P(S e) > b -, P (S e) > b, where S is the sentence in question, e is the current knowledge (evidence), b is the break.even probability determined from the payoff. The system has all the desired properties (if the proof theory of its probability logic is sound), so provides insight into nonmonotonic reasoning, an important facet of intelligent behavior. Acknowledgments This work was supported in part by the Office ofn a val Research under grantnooo K Thanks also to Larry Rendell, Peter Cheeseman, Marianne Winslett for their valuable comments on an early version of this paper. 203

8 REFERENCES Cheeseman, P. (1985). n defense of probability. n Proc./JCAJ-85. Dowty, D.R., Wall, R.E., & Peters, S. (1981). ntroduction to Montague Semantics. Dordrecht, Holl: Deidel Publishing. Doyle, J. (1983). A truth maintenance system. Artificial ntelligence, 12, Dubois, D. &Prade, H. (1988). Default reasoning possibilitytheory.artificiallntelligence, 35, Farreny, H. & Prade, H. (1986). Default inexact reasoning with possibility degrees. Tran. on Systems, Man, Cybernetics, 16, no. 2., Ferguson, T.S. (1967). Mathematical Statistics: A Decision Theoretic Approach. New York: Academic Press. Frisch, A.M. & Haddawy, P. (forthcoming). Probabilistic as a modal operator. Proc. of the Fourth Workshop on Uncertainty in Artificial ntelligence. Ginsberg, M.L. (1984). Non-monotonic reasoning usj.ng Dempster's rule. n Proc. AAA-84 (pp ). Ginsberg, M.L. (1985). Does probability have a place in non-monotonic reasoning? n Proc. /JCA-85 (pp ). Ginsberg, M.L. (1987). ntroduction of Readings in Nonmonotonic Reasoning, Los Altos, CA: Morgan Kaufmann Publishers. Kyburg, H.E. & Smolder, H.E. (1964). ntroduction of Studies in Subjective Probability. New York: John Wiley & Sons. Kyburg, H.E. (1970). Conjuctivitis. n M. Swain (Ed.), nduction, Acceptance, Rational Belief. Reidel. Levi,. (1980). The Enterprise of Knowledge. Cambridge, MA: MT Press. Lifschitz, V. (1987). Pointwise Circumscription (Technical Report). Stanford, CA: Stanford, University. Reprinted in M.L. Ginsberg (Ed.), Readings in Nonmonotonic Reasoning. Los Altos, CA: Morgan Kaufmann Publishers. McCarthy, J. & Hayes, P. (1969). Some philosophical problems from the stpoint of artificial intelligence. Machine ntelligence 4. Reprinted in B. Webber & N. Nilsson (Ed.), Readings in Artificial ntelligence. Tioga, McCarthy, J. (1980). Circumscription-- A fonn of non-monotonic reasoning. Artificial ntelligence, 13, McDermott, D. (1982). Non-monotonic logic. Journal ACM, 29, Moore, R.C. (1985). Semantical considerations on nonmonotonic logic. Artificial ntelligence, 25, Nilsson, N. (1986). Probabilistic logic. Artificial ntelligence, 28. Pearl, J. (1987). Distributed revision of composite beliefs. Artificial/ ntelligence, 33, Reiter, R. & Criscuolo, G. (1981). On interacting defaults. Proc. /JCA-81 (pp ). Reiter, R. ( 1980). A Logic for Default Reasoning. Artificial ntelligence, 13, Rich, E. (1983). Default reasoning as likelihood reasoning. Proc. AAA-83 (pp ). Savage, L.J. (1954). The Foundations of Statistics. John Wiley & Sons. Yager, R.R. (1987). Using approximate reasoning to represent default knowledge. Artificial ntelligence, 31,